Optimal noise in spiking neural networks for the detection of chemicals by simulated agents

We created a spiking neural controller for an agent that could use two different types of information encoding strategies depending on the level of chemical concentration present in the environment. The first goal of this research was to create a simulated agent that could react and stay within a region where there were two different overlapping chemicals having uniform concentrations. The agent was controlled by a spiking neural network that encoded sensory information using temporal coincidence of incoming spikes when the level of chemical concentration was low, and as firing rates at high level of concentration. With this architecture, we could study synchronization of firing in a simple manner and see its effect on the agent’s behaviour. The next experiment we did was to use a more realistic model by having an environment composed of concentration gradients and by adding input current noise to all neurons. We used a realistic model of diffusive noise and showed that it could improve the agent’s behaviour if used within a certain range. Therefore, an agent with neuronal noise was better able to stay within the chemical concentration than an agent without.